[dm-devel] limits->max_sectors is getting set to 0, why/where? [was: Re: dm: kernel oops by divide error on v4.16+]

[Kees Cook]

On Mon, Apr 9, 2018 at 2:56 PM, Jens Axboe <axboe at kernel.dk> wrote:
> On 4/9/18 3:26 PM, Jens Axboe wrote:
>> On 4/9/18 1:32 PM, Jens Axboe wrote:
>>> On 4/9/18 12:38 PM, Mike Snitzer wrote:
>>>> On Mon, Apr 09 2018 at 11:51am -0400,
>>>> Mike Snitzer <snitzer at redhat.com> wrote:
>>>>
>>>>> On Sun, Apr 08 2018 at 12:00am -0400,
>>>>> Ming Lei <ming.lei at redhat.com> wrote:
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> The following kernel oops(divide error) is triggered when running
>>>>>> xfstest(generic/347) on ext4.
>>>>>>
>>>>>> [  442.632954] run fstests generic/347 at 2018-04-07 18:06:44
>>>>>> [  443.839480] divide error: 0000 [#1] PREEMPT SMP PTI
>>>>>> [  443.840201] Dumping ftrace buffer:
>>>>>> [  443.840692]    (ftrace buffer empty)
>>>> ...
>>>>>> [  443.845756] CPU: 1 PID: 29607 Comm: dmsetup Not tainted 4.16.0_f605ba97fb80_master+ #1
>>>>>> [  443.846968] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.10.2-2.fc27 04/01/2014
>>>>>> [  443.848147] RIP: 0010:pool_io_hints+0x77/0x153 [dm_thin_pool]
>>>>
>>>> ...
>>>>
>>>>> I was able to reproduce (in my case RIP was pool_io_hints+0x45)
>>>>>
>>>>> Which on my kernel, is:
>>>>>
>>>>> crash> dis -l pool_io_hints+0x45
>>>>> /root/snitm/git/linux/drivers/md/dm-thin.c: 2748
>>>>> 0xffffffffc0765165 <pool_io_hints+69>:  div    %rdi
>>>>>
>>>>> Which is drivers/md/dm-thin.c:is_factor()'s return
>>>>> !sector_div(block_size, n);
>>>>>
>>>>> SO looking at pool_io_hints() it would seem limits->max_sectors is 0 for
>>>>> this xfstests device... why would that be!?
>>>>>
>>>>> Clearly pool_io_hints() could stand to be more defensive with a
>>>>> !limits->max_sectors negative check but is it ever really valid for
>>>>> max_sectors to be 0?
>>>>>
>>>>> Pretty sure the ultimate bug is outside DM (but not seeing an obvious
>>>>> place where block core would set max_sectors to 0, all blk-settings.c
>>>>> uses min_not_zero(), etc).
>>>>
>>>> I successfully ran this test against the linux-dm.git
>>>> "for-4.17/dm-changes" tag that Linus merged after the block changes:
>>>>  git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git tags/for-4.17/dm-changes
>>>>
>>>> # ./check tests/generic/347
>>>> FSTYP         -- ext4
>>>> PLATFORM      -- Linux/x86_64 thegoat 4.16.0-rc5.snitm
>>>> MKFS_OPTIONS  -- /dev/mapper/test-xfstests_scratch
>>>> MOUNT_OPTIONS -- -o acl,user_xattr /dev/mapper/test-xfstests_scratch /scratch
>>>>
>>>> generic/347      65s
>>>> Ran: generic/347
>>>> Passed all 1 tests
>>>>
>>>> SO this would seem to implicate some regression in the 4.17 block layer
>>>> changes.
>>>
>>> No immediate ideas come to mind, we didn't have a lot of changes and I
>>> don't see anything that looks problematic. Maybe you can try and
>>> bisect it and see what you come up with?
>>
>> I ran it, problematic commit is:
>>
>> commit 3c8ba0d61d04ced9f8d9ff93977995a9e4e96e91
>> Author: Kees Cook <keescook at chromium.org>
>> Date:   Fri Mar 30 18:52:36 2018 -0700
>>
>>     kernel.h: Retain constant expression output for max()/min()
>>
>
> The fun continues. Thinking I'd try a userspace repro and thinking it
> would be difficult to reproduce, try the attached min.c that just copies
> all the bits from include/linux/kernel.h
>
> axboe at x1:~ $ gcc -Wall -O2 -o min min.c
> axboe at x1:~ $ ./min 128 256
> min_not_zero(128, 256) = 0

This should be fixed with e9092d0d9796 ("Fix subtle macro variable
shadowing in min_not_zero()").

-Kees

-- 
Kees Cook
Pixel Security